prof. Ing. Vanda Benešová, CSc.

Computer vision is gaining an increasingly important position in the automatic processing of medical visual data, especially radiological and histological images. The most important current topics of computer vision research in medical applications are related to the diagnosis of various diseases and their goal is to provide the doctor with additional relevant information, or to relieve him of some tasks in an automatic or semi-automatic mode.

To meet these goals, the development of new, robust computer vision methods using the modern approaches of deep neural networks is necessary.

We see challenges not only in the research of new methods of computer vision using deep learning, research of their interpretability and explainability, in the generation of data for the data augmentation, but also in the optimization of the process of iterative development of a new medical application, including the efficiency of the annotation process.

Research during the doctoral studies will focus on one of the mentioned areas.

prof. Ing. Vanda Benešová, CSc.

prof. RNDr. Pavel Surynek, Ph.D.

Department of Applied Mathematics

This thesis deals with lesion segmentation of hypoxic-ischemic encephalopathy in neonatal MRI images using deep neural networks. The work explores and implements various approaches such as super resolution and data synthesis to achieve more accurate segmentation on the BONBID-HIE dataset. As part of this work, we implemented a functional pipeline for creating super resolution 3D MRI images, a pipeline for creating synthetic lesions that were further inpainted into the dataset images, and a segmentation pipeline. The results were discussed and compared.

prof. Ing. Vanda Benešová, CSc.

Ing. Daniel Vašata, Ph.D.

Department of Applied Mathematics

This master's thesis focuses on advanced methods for processing digital histopathological images to improve melanoma cancer prediction using the PUMA dataset. The work focuses on the analysis of current solutions for panoptic segmentation in histopathological data. It proposes a segmentation pipeline in the form of the selected state-of-the-art model TransUnet that is further modified, followed by an Autoencoder for stitching the image patches. This pipeline overcomes the baseline by a 0.06 average DICE score. The thesis includes a detailed description of data preprocessing, hyperparameter optimization, and the implementation of selected models.